Reproduce “Improving Uncertainty Estimation through Semantically Diverse Language Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lukas Aichberger, Kajetan Schweighofer, Mykyta Ielanskyi, Sepp Hochreiter
Semantically Diverse Language Generation (SDLG) estimates uncertainty in language-model answers by actively generating meaningfully different alternatives rather than relying on random samples alone. Across OPT models and three question-answering datasets, SDLG improved uncertainty-based correctness detection—most notably on TruthfulQA, where AUROC reached up to 0.956—and generated more semantic diversity while reducing token-generation costs by 15–33%.
Key insights
- Semantic uncertainty should be computed over distributions of meanings rather than over individual token sequences, since different wordings may express the same answer.
- Actively seeking semantic changes produces substantially more distinct answer clusters than multinomial sampling; on CoQA, SDLG increased the number of clusters by 19% after two responses and 74% after ten.
- The method estimates aleatoric uncertainty in a model’s own predictive distribution, so it can identify ambiguity but cannot by itself detect facts that the model confidently believes incorrectly.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.